DRAFT – Do Not Cite Without Permission Factors Supporting Success in Return-To-Work Programs for Persons with Severe Disabilities: An Exploratory Analysis from Wisconsin’s SPI Project Barry S. Delin Stout Vocational Rehabilitation Institute University of Wisconsin – Stout Anne E. Reither Utah State University November 2006 Paper prepared for the Association of Public Policy Analysis and Management Research Conference, Madison, WI, November 2-4, 2006. This paper builds on research performed at the University of Wisconsin – Stout Vocational Rehabilitation Institute and the Oregon Health and Sciences University Health Policy Institute on behalf of the Pathways Projects, Office of Independence and Employment, Wisconsin Department of Health and Family Services. The authors thank the staff at the Pathways Projects for their cooperation and support. The descriptions and interpretations in this paper are those of the authors and do not reflect those of the Pathways Projects, the Wisconsin Department of Health and Family Services, or of the institutions where the authors are employed. DRAFT – Do Not Cite Without Permission 1 Introduction From July 1999 through September 2004, the Wisconsin Department of Health and Family Services (DHFS) operated a “return-to-work” demonstration project for disabled Social Security beneficiaries and recipients as part of the State Partnership Initiative. This demonstration was called Wisconsin Pathways to Independence (WPTI) and was one of twelve State Partnership Initiative (SPI) projects funded by the Social Security Administration (SSA) through co-operative agreements with the participating states. Under these co-operative agreements states were given the opportunity to develop and implement new or augmented programs intended to improve employment outcomes for persons receiving Social Security Disability Insurance (SSDI) and/or Supplemental Security Income (SSI). There was an expectation that the SPI demonstration projects would also provide the states, SSA, and other federal entities with information that could inform future program planning and policy development. 1 WPTI resulted in what can be characterized as modest, but significant, gains in the probability of employment, in earnings, and in income for those who received the intervention. At the time of study entry participants in the intervention had somewhat lower employment rates and quarterly earnings than those in the comparison group. 2 Regression models of program effects showed that participants in the intervention group exhibited a gain of 13% in their employment rate relative to the comparison group one year after study entry. This difference remained at 12% two years following study entry. Average gains in earnings were estimated at $216 per calendar quarter after one year, $314 per quarter after two. Those receiving the intervention achieved an estimated income growth of $314 per calendar quarter relative to that estimated for comparison 1 The Rehabilitation Services Administration (RSA), U.S. Department of Education also sponsored a number of SPI projects. The RSA sponsored projects were not restricted to Social Security beneficiaries and recipients and were not required to collect as extensive a range of data as the SSA sponsored projects. 2 Delin, Barry S., Reither, Anne E., Drew, Julia A. and Hanes, Pamela P. 2004. Final Project Report: Wisconsin Pathways to Independence. Menomonie, WI: University of Wisconsin – Stout Vocational Rehabilitation Institute, pp. 93-97 and 99-101. These employment outcomes were measured using Unemployment Insurance (UI) data from the Wisconsin Department of Workforce Development. WPTI did not use random assignment. The comparison group was recruited from Wisconsin Division of Vocational Rehabilitation (DVR) consumers who appeared to meet WPTI eligibility requirements. Additionally, the comparison group was limited to those DVR consumers who had progressed far enough through the vocational rehabilitation process to provide behavioral evidence of strong motivation to become employed or to increase earnings. There were significant differences between the two groups in the distributions of some measured baseline characteristics, including employment and earnings. We believe differences in how the two groups were recruited and the typical timing of entry to the study explained much of the observed differences. Comparison group members were typically well into and had sometimes completed their latest span of DVR services. By contrast, intervention group members were typically starting an initial or new span of DVR services and may in some cases have reduced their work efforts in anticipation of seeking employment services. The original study utilized analysis techniques to adjust for the baseline differences between groups. DRAFT – Do Not Cite Without Permission 2 group members. The estimated difference per quarter was $456 two years after study entry. 3 The WPTI evaluation emphasized average differences between the intervention and comparison groups. This was in keeping with SSA’s first priority: learning what the typical effects of the various SPI initiatives were over the populations defined by the projects’ eligibility criteria and recruitment processes. 4 However, this paper does not focus on differences between individuals in the intervention and comparison groups. The point of departure is the fact that there was substantial variation in intervention group outcomes. There was not a single calendar quarter in which a majority of those in the WPTI intervention group (or, for that matter, the comparison group) had employment or earnings reported in Wisconsin Unemployment Insurance records. This finding was not surprising for a group of individuals who were all SSDI beneficiaries and/or SSI recipients for reason of disability. After all, every one of these individuals had to demonstrate that they met the Social Security definition of disability, a definition that requires that an individual cannot work at any job in the economy that can generate monthly earnings at or above the Substantial Gainful Activity (SGA) level. It would seem useful to learn the sources of such variation in order to improve return-to-work programs and policies and/or to identify those individuals more likely to benefit from those program and policies. Indeed, under current law, any decision by a beneficiary or recipient to seek employment is ultimately voluntary. This exploratory study utilizes logistic regression to attempt to identify the reasons that some WPTI participants (the term “participant” should from this point forward be understood as referring only to those in the intervention group) had much better employment outcomes than others. The WPTI Intervention Wisconsin Pathways to Independence was operated jointly by what is now the Office of Employment and Independence (OIE) in the Department of Health and Family Services and, until October 2003, the Division of Vocational Rehabilitation (DVR) in the Department of Workforce Development. The project was intended to address a broad range of issues related to service delivery and the negative work incentives embedded in public policies that WPTI’s designers and proponents believed had negatively affected the ability of those in the SSDI and SSI programs to become employed, remain employed, increase their earnings or, if desired, engage in career development. Consequently, WPTI involved features to integrate and improve service provision and to 3 Delin, Barry S., et. al. 2004. pp. xiv-xv. Also see chapter IV of that report, “Net Effects of WPTI,” pp. 131-79. The actual observed differences between the intervention and comparison groups (as opposed to estimated differences) were smaller for employment (8% for both years), but of somewhat larger size for earnings and income. Income was approximated by the sum of earnings and social security benefits, including a SSI state supplement. For example, the earnings difference per quarter was $306 per quarter one year after study entry, $364 after two. All monetary values were converted to constant dollars using 1996 GDP = 100. 4 SSA was initially interested in identifying ways of increasing employment and earnings for those on SSDI and SSI in hope that some would be able to permanently exit the programs. Over time, perhaps in part from what was learned through SPI, SSA placed increasing emphasis on having beneficiaries and recipients reduce their dependency on SSA benefits without necessarily leaving benefit status. DRAFT – Do Not Cite Without Permission 3 implement, whether on a temporary or permanent basis, changes to public policy. The intent was to create a process that would maximize the participant’s ability to overcome whatever barriers impeded the achievement of the participant’s employment goals. Existing deficiencies in service provision and policy were seen as having subjective as well as objective effects; individuals might refrain from increased work effort because of their fears, justified or not, about the negative consequences. WPTI provided participants with a combination of “work incentive” benefits counseling and person centered vocational planning services. Participants were enrolled at and served by twenty, largely non-profit, community agencies. Enrollment at each agency was restricted to consumers deemed to fit into a particular disability category. Consequently, WPTI had two kinds of eligibility requirements. One type pertained to overall project eligibility, the second type constrained entry to the project through specific types of community agencies. Participants had to be in either the SSDI or SSI program (for reason of disability), Wisconsin residents, between eighteen and sixty-four years of age (inclusive) and eligible for DVR services. 5 WPTI designated four types of community agencies for purposes of the project, each of which served persons with a disabling condition that fit under one of four categories, physical disabilities, mental health, developmental disabilities, and HIV/AIDS, that were defined based on categories already used by DHFS. 6 If a potential participant met eligibility criteria, entry to the program was by selfselection, modified by potentially three considerations: the community agency had to have the capacity to serve the individual, DVR had to be willing to authorize the participant to receive WPTI services, and in some cases the participant had to meet legal requirements for receiving services from a particular community agency. 7 The probability that any individual would enroll in WPTI was greatly increased if that person was a current or recent consumer at one of the community agencies and/or DVR. Additionally, the proportion of participants at each of the four types of provider agencies closely mirrored the availability of program slots. Thus, it isn’t surprising that the distribution of disabilities and some socio-demographic characteristics of WPTI participants diverged substantially from those exhibited by the comparison group. Finally, there were two periods during the WPTI demonstration when, due to fiscal 5 Nine participants were allowed to enroll without DVR authorization. These individuals were persons who had had AIDS or a disabling condition related to their HIV positive status and did not wish to reveal their condition to DVR. Several sites were authorized to enroll persons sixteen and seventeen years old who were engaged in transition activities. Only one such participant is included in the 506 cases utilized in this paper. 6 Though a community agency was only allowed to enroll those consumers who met the WPTI definition for these groups, many consumers met more than one of these definitions. In those regions of the state where there were multiple WPTI associated community agencies, a consumer potentially had some choice as to where to enroll and thus to the style of person centered vocational planning the consumer would receive. 7 At certain agencies, entrants to WPTI had to become clients of the agency first. In some cases this reflected agency philosophy, but generally this reflected a combination of state regulations and limited public funding. This was particularly true in the case of the Community Support Programs which serve persons with severe and persistent mental illness. DRAFT – Do Not Cite Without Permission 4 shortfalls, DVR closed or limited access to its services. These Order of Selection closures both slowed enrollment and affected the timing and level of service provision. Benefits counseling was implemented using a single model. There were, however, multiple models used to provide vocational planning services. Community agencies were required to use a vocational planning approach the WPTI central office had approved for use with consumers from the disability category each agency was authorized to enroll. 8 The delivery of benefits counseling and vocational planning services was supposed to be integrated through a consumer directed team process involving the participant, community agency staff, the DVR counselor, and other persons agreeable to the participant. Particular attention was given to conducting an in-depth assessment of barriers that made achievement of employment goals difficult, followed by development of a plan to identify and mobilize necessary resources to deal with identified barriers. There was an acceptance that there could be substantial variation in the delivery of the WPTI service approach based upon participant needs and circumstances. In fact, this personalized approach was a key part of the intervention. All participants, in principle, had access to the combination of benefits counseling and vocational planning services. Additionally, it was expected that the consumer directed team, especially the community agency staff and DVR counselor, would work to help the participant gain or maintain access to services and supports that WPTI did not itself provide. Although there was an expectation that most participants would receive the largest amount of WPTI services during their first months of participation, it was also expected that WPTI would continue to provide services and supports “as needed” until the end of the demonstration. 9 WPTI was not intended as a means to achieve quick placement into gainful employment. The designers believed that better long term outcomes would be achieved through a vocational planning process that allowed participants to identify goals and how to achieve them. So it was up to the participant and her/his team to decide the best time to (re)enter the workforce. 10 8 The Physical Disability and HIV/AIDS agencies used the same vocational planning model, the Vocational Futures Planning model that was originally intended to be the WPTI standard. The Developmental Disabilities agencies were suppose to use a second model, but were in fact given broad latitude to deliver whatever model they chose. The Mental Health sites used one of two other models, depending on whether the agency was a Community Support Program or a Clubhouse. Although all models were said to be team based and consumer centered, there was strong evidence of variation in how agencies using the “same” model interpreted how that model should be delivered. See Delin, Barry S., et al. 2004. pp. 17-23. 9 Only about 45% of benefits counseling and vocational services (excluding job coaching and various forms of indirect support) were delivered in the first six to nine months of participation (i.e., in the first two calendar quarters after the enrollment quarter and whatever portion of the enrollment quarter followed the actual enrollment date). There were both programmatic (e.g. staff attrition, DVR budget shortfalls) and participant (e.g., illness, reluctance to commit to the process) reasons why an unexpectedly small proportion of service hours were delivered in the first months following enrollment. See Delin, Barry S., et al. 2004. p. 110. 10 This proved to be an area of some tension between DVR staff and both DHFS and agency staff. DVR as an agency and DVR area directors and counselors as individuals are evaluated on the number/proportion of successful closures. In a context of insufficient funds and waiting lists, DVR personnel had understandable concerns about the length and cost of the intervention. DRAFT – Do Not Cite Without Permission 5 Additionally, WPTI sought and to some extent obtained changes to health care and income support program provisions that could potentially support better employment outcomes for WPTI participants. These policy changes were only applicable to a subset of WPTI participants, though program designers initially expected these subsets to be fairly large. WPTI was instrumental in getting Wisconsin to adopt a Medicaid Buy-in program for working people with disabilities, abbreviated as MAPP. As this program was created through statute, MAPP was available to any Wisconsin resident who met the eligibility requirements. However, WPTI’s designers had sought MAPP in order to provide access to public health care coverage for WPTI participants, especially SSDI beneficiaries, who might be more willing to take their earnings “permanently” high enough to leave SSA benefits if access to a public health care program might be preserved. 11 SSA also approved a less severe earnings offset for SSI recipients in WPTI. This “SSI waiver” allowed eligible participants to keep three dollars of their SSI cash benefit for every four dollars of earnings and allowed participants to have savings not otherwise allowable under SSI and Medicaid. 12 WPTI also asked SSA for permission to test an offset for SSDI beneficiaries to address the severe disincentive to earnings over SGA faced after the end of the Trial Work Period. SSA did not approve this request. 13 WPTI also planned to seek waivers from provisions of the Foods Stamp and Section 8 Housing programs, but quickly abandoned these efforts to concentrate on obtaining waivers from SSA. Ultimately, 956 individuals enrolled in WPTI. However, only 506 of these participants were included in the outcomes evaluation completed December 2004, largely because the excluded cases did not have the necessary two years of data following their entry into the project that was required for inclusion in the analysis. These 506 participants constitute the sample that is used for the analyses presented in this paper. 14 11 MAPP, as a Balanced Budget Act buy-in, is not restricted to persons under full Social Security retirement age. Participants must meet the Social Security disability standard, though having earnings at or above the SGA level does not preclude eligibility. While employment or participation in a program leading to employment is required, there is no specific requirement that participants have monetary earnings. Even though subsequent changes to federal law allowed former SSDI beneficiaries to retain Medicare eligibility longer than the end of the Extended Period of Eligibility, Medicaid (especially prior to Medicare Part D) provided a range of services and support generally not available through Medicare. 12 The 1619 provision of the SSI includes a 1:2 offset and continued access to Medicaid after the cash benefit is “zeroed out.” However, 1619 does not provide relief from Medicaid asset limits. 13 However, in 2005, Wisconsin became one of four states selected to pilot a SSDI earnings offset in preparation for a national demonstration. 14 The 506 includes participants who enrolled through September 30, 2001. It would appear, in theory, that this number could be augmented by those who enrolled through September 30 of the following year. However, survey data is available only for those who enrolled by December 31, 2001; i.e. another three months. We decided that the work involved in augmenting an available data set to gain about another twenty-five cases was not justified in the context of this exploratory study. DRAFT – Do Not Cite Without Permission 6 Conceptual Approach Wisconsin Pathways to Independence was intended, above all else, to test an approach to increasing the employment rates and earnings of working age persons attached to a Social Security disability program. Unemployment Insurance (UI) data was the principal source of employment and earnings data utilized to assess participant outcomes for the WPTI evaluation. UI data are available on a calendar quarter basis. The data set used for the WPTI evaluation was organized into a seventeen calendar quarter time series with the calendar quarter of enrollment preceded and followed by eight calendar quarters of data. 15 Though participants in Wisconsin Pathways to Independence achieved significant gains in their employment rate and average earnings relative to members of the comparison group, the stark fact remains that there was not a single calendar quarter during the entire seventeen quarter period when their employment rate calculated from UI records reached 50%. In the calendar quarter immediately prior to the enrollment quarter only 35% of future WPTI participants were employed. The maximum employment rate of 46% was reached in quarters four and five following the enrollment quarter, followed by a fairly rapid decline to 40% at quarter eight. 16 These results were not completely unexpected. Though Social Security programs include work incentives, SSDI beneficiaries and SSI recipients need to meet the Social Security disability definition to retain attachment to the programs. Whatever the variation may be in the actual application of the definition, the definition for adults remains strict: the inability to engage in work that is compensated at the substantial gainful activity (SGA) level due to a physical or mental impairment that has lasted or is expected to last at least one year or to result in death. Second, as previously noted, SSDI beneficiaries or SSI recipients face no legal compulsion to attempt paid employment as long as they maintain their disability status. At the same time, those on SSDI and SSI always face the possibility that work activity can serve as evidence of medical improvement. Moreover, should they still decide to work, many persons on Social Security disability programs can face significant declines in income or losses of important non-cash benefits, despite the availability of work incentives. Given the time and effort it took many SSDI beneficiaries and SSI recipients to establish their eligibility for Social Security and other public benefits, it should not be surprising that many persons capable of and/or desiring to perform some work do not. According to a frequently cited statistic from the (former) 15 Wisconsin UI data does not report all employment. Most notably self-employment and jobs at out of state employers are excluded. However, in addition to this underreporting, there is a way in which the UI data may overestimate the employment rate. The UI records do not include direct information about how much time within the quarter an individual is employed, and therefore may include even very brief periods of work activity. 16 The employment rate had its minimum value (32%) four quarters prior to the enrollment quarter. Thereafter, approaching the enrollment quarter, the rate exhibited a slow rising trend, seemingly inconsistent with the expectation that persons tend to leave work just before entering employment programs. Though not appearing in the WPTI final report, this material has been presented at several public events. For example, see Delin, Barry S., Reither, Anne E., and Drew, Julia A. 2003. “Employment and Earnings Trends in a Complex Program Delivery Environment.” Presentation at the State Partnership Initiative Annual Meeting, Washington D.C., August 7-8, 2003. DRAFT – Do Not Cite Without Permission 7 General Accounting Office, less than 1% of those participating in a Social Security disability program ever leave the rolls to return to work. 17 Still, the proportion of those attached to Social Security disability programs who have some reported earnings can be much higher than implied by the GAO number. For example, the proportion of working age Wisconsin SSI recipients performing some paid work during the period WPTI was operating ranged between 17% and 21%. 18 We believe that one of the lessons learned, or at least reinforced, through WPTI and the State Partnership Initiative is that there can be value to those with disabilities, SSA and other government entities, and the wider community in increasing employment and earnings and reducing dependency on benefits, even when that does not result in individuals ending their attachment to a Social Security disability program. Thus, even when a return-to-work program does not have a large impact on average employment rates and earnings across the SSA disability program population, that program may constitute good policy if one can both identify the attributes and conditions that make it more likely for some to benefit and then identify and encourage participation in return-towork programs by such persons. In our search for a method to identify the characteristics of those who had benefited most from their WPTI participation, we found a study conducted by Chi-Fang Wu, Maria Cancian, and Daniel R. Meyer that approached this issue in the context of changes in earnings and employment of TANF participants. 19 While there are many differences between the TANF and Social Security disability populations and the legal requirements and political contexts in which the respective programs operate, there are commonalities in that both populations can be viewed as being composed of individuals who find it difficult to become employed, remain employed, and/or increase earnings. Wu, et al. sought to identify TANF participants who achieved what they characterized as better medium-term and long-term employment outcomes. A key step in doing this was to categorize TANF participants into earnings and employment patterns or trajectories associated with different levels of success over 17 Although this number has been used for both Social Security disability programs, its origin appears to be as an estimate of the rate of departure from SSDI because of work. See General Accounting Office. March 1997. Social Security Disability Programs Lag in Promoting Return to Work. (GAO/HEHS-97-46) Washington, DC: General Accounting Office, p.1. Also, there have been estimates of somewhat higher departure rates based upon the analyses of specific cohorts of SSDI entrants. In the most optimistic case (a study of entrants under age 40), about 4% left the program because of return to work. See Mashaw, Jerry L. and Reno, Virginia P., eds. 1996. The Challenge of Disability Income Policy. Washington, DC: National Academy of Social Insurance. pp. 109-11. 18 These data were obtained from relevant issues of the “SSI Disabled Recipients Who Work” report and have been adjusted to reflect only recipients who were age 18-64. During the same period the national rate was estimated as being in the 8% to 9% range. 19 Wu, Chi-Fang, Cancian, Maria, and Meyer, Daniel R. 2005. “Standing Still or Moving Up? Evidence from Wisconsin on the Long-Term Employment and Earnings of TANF Participants” Unpublished Manuscript presented at the Association of Public Policy Analysis and Management Research Conference, Washington D.C., November 2005. Wu, et al. have graciously given permission to cite this draft version of their work. DRAFT – Do Not Cite Without Permission 8 some defined time period following TANF entry. Both earnings and employment trajectories were defined in terms of the level of change, its directionality, and the stability of trends; each case was assigned to a separate earnings and employment trajectory. 20 Once participants were assigned to a trajectory, the authors were able to compare the distribution of characteristics and experiences across these groupings. Wu, et al. utilized both descriptive methods and multinomial logistic regression models to analyze their data. Medium-term outcomes were defined as those occurring three years after entering TANF; long-term outcomes were those occurring six years following entry. As will become apparent, we borrowed heavily from the conceptual approach developed by Wu, Cancian, and Meyer. Like Wu, et al., we began by defining earnings and employment trajectories. However we did so facing some limits imposed by the data available from WPTI. The most important was the limited number of cases, 506 compared to the approximately 17,000 available for the TANF study. Consequently, we defined substantially fewer trajectories for each outcome variable and then defined dichotomous categories of relative success for each variable. We present evidence below that supports the appropriateness of these choices. Also, we had data for a period of time that Wu, et al. would characterize as medium-term. The final calendar quarter of data available was for the eighth quarter following the calendar quarter of WPTI enrollment. As we had access to Unemployment Insurance earnings and employment data for a period prior to enrollment, we included the quarter previous to the enrollment quarter so that the initial data period does not include any delivery of WPTI services. 21 Thus, the data period for these outcome variables is ten calendar quarters or two and one-half years. We also constructed an alternative employment “success” indicator variable using participant reported data collected on a monthly basis by staff at the community agencies. These data were available only from the enrollment quarter forward, limiting the analysis period to nine quarters. 22 These limited data spans could have an important consequence. Wu, et al. observed that in their study population a desirable medium-term trajectory did not consistently lead to good long-term outcomes. Given the cyclical nature of some 20 See Wu, et al. 2005. pp. 5-7. 21 We did not use UI data from before Q-1 as we are only interested in changes in outcomes following WPTI enrollment. Additionally, we would not have encounter and survey data for that earlier period. 22 Depending on the participant’s enrollment date, the period of WPTI participation can vary from a minimum of twenty-four full months to a maximum of twenty-seven. Additionally, we had originally hoped to be able to analyze outcomes over a much longer period for a significant minority of our cases. Wisconsin is one of four states currently piloting an earnings offset for the SSDI program. Program planners anticipated that approximately half of the 800 participants would be former WPTI participants, as many of the same community agencies would be involved in enrollment and the provision of benefits counseling to pilot participants. The former WPTI participants are asked to sign an additional release allowing researchers to use data collected during WPTI to the new data collected for the pilot project. Unfortunately, only 5% of pilot participants (through August 2006) were former WPTI participants. DRAFT – Do Not Cite Without Permission 9 disabilities, the progressive nature of others, and the time limited nature of some Social Security work incentives, this is an important caution. 23 Outcome Measures: What Constitutes Success? Three dichotomous outcome measures were constructed to summarize participant earnings and employment outcomes for each of the 506 participants. Data from Wisconsin unemployment Insurance (UI) records were used to construct indicators of relative success in the areas of earnings and employment. Data from monthly participant update forms were used to construct an alternative indicator of employment success which would include employment not captured in the UI system. 24 The UI based indicators utilize ten calendar quarters of data, the alternative form-based employment indicator, for reasons identified above, utilizes nine quarters of data. Each indicator was constructed from two types of underlying trajectories, one in the area of growth, the other attempting to capture persistence in achieving what was defined as a “good” outcome. Both growth and persistence were assessed by comparing a base year to an outcome year. 25 While this approach has the undesirable property of ignoring quarter to quarter variation in any individual’s outcomes, we think accepting this limitation has the virtue of being consistent with the fact that many people with disabilities will leave work for relatively short periods for reasons not under their direct control such as variations in the symptomatic severity of their disabling condition or the loss of important public program or natural supports. We sought to define “success” in ways that would differentiate participants without a large number near the boundaries between the categories. We also sought to define success at levels that would insure enough cases in the “more successful” group to allow analysis. Given the limited number of cases we chose to forgo analysis of what might be characterized as the most successful participants of all (for example, those with earnings more than twice the poverty level). In the case of the UI earnings indicator, growth was defined as an increase of more than the mean increase between the base and the outcome year ($949 in 1996 GDP dollars). 26 Persistence was defined as 23 Wu, et al. 2005. p. 12. 24 However, the employment data reported from the community agencies contains some inaccuracies. That data did not always include jobs reported in the UI system. See Delin, et al. 2004. p.141. 25 For both UI based indicators the base year is Q-1 through Q2 (where Q0 designates the enrollment quarter). For the Forms Employment Indicator the base year is Q0 through Q3. We also constructed UI indicators using Q0 through Q3 as the base year. Their distributions proved almost identical to those for the Q-1 through Q2 base years. We did some preliminary modeling with these variables and found they performed similarly to the Q-1 through Q2 formulations. The more serious problem is the relatively short duration between the base and reference years which compromises our claim to be measuring medium-term, as opposed to short-term change. 26 The median change in UI earnings between the base and outcome years was $0 reflecting the fact that a large number of participants had no UI earnings reported in either the base or outcome year. The median value was nearly one quarter standard deviation below the mean value. There is a significant tail representing cases where there was a decrease in earnings. DRAFT – Do Not Cite Without Permission 10 maintaining earnings above $2500 per year in both the base and outcome year. This seemingly arbitrary figure roughly captures the annualized trigger amount for a SSDI Trial Work Period month when Pathways began in 1999. 27 There was a moderate tendency for a participant who was identified as “more successful” in earnings persistence to also be identified as being in the higher earnings growth category. 28 For both the UI and update form based indicators, growth was defined as moving from having either zero, one, or two quarters in which employment was reported during the base year, to having employment reported in either three or four quarters in the outcome year. 29 Success in persistence meant having either three or four quarters in which employment was reported in both the base and outcome years. 30 Table 1 shows that the success indicators meet the design criteria of capturing a reasonably high number of the 506 cases in each of the two categories. Do they adequately differentiate relatively successful cases from less successful ones? Table 1: Earnings and Employment Success Indicators (Numbers and Percentage per Category) More Successful Less Successful UI Earnings Success 236 (46.6%) 270 (53.4%) Indicator (UI$I) UI Employment Success 174 (34.4%) 332 (65.6%) Indicator (UIEI) Form Employment Success 226 (44.7%) 280 (55.3%) Indicator (FEI) The employment and earnings trends displayed, respectively, in Figures 1 and 2 are consistent with an interpretation that the “more successful” and “less successful” values of each dichotomous indicator variable represent groups with substantially 27 SSDI beneficiaries, including those with concurrent SSI eligibility are allowed to test their ability to work during a nine month Trial Work Period (TWP). During the TWP, the beneficiary retains all of their cash benefit. The months do not need to be consecutive, but months where earnings are above a certain monthly amount count as one of the nine months. During the thirty-six month Extended Period of Eligibility that follows completion of the TWP, beneficiaries who retain their eligibility following a medical review and earn at or above SGA in any month will lose their cash benefit for that month. Those who have earnings less than SGA will receive their full SSDI cash benefit. 28 About 85% of those in the higher growth category were also in the higher persistence category. The reciprocal association was much weaker. Only about 55% of those classified into the higher persistence category were also classified into the higher growth category. 29 Because of the manner in which UI Employment is reported, it is impossible to know what time period an individual was actually employed in a calendar quarter. It may have been the full three months; it may have been a single day. We used the same convention for the participant reported information, although these data could, in principle, be used to calculate the proportion of time in a calendar quarter an individual actually was employed. 30 For the employment indictors, the persistence and growth criteria are, by definition, logically exclusive. DRAFT – Do Not Cite Without Permission 11 different and diverging outcomes following entry to WPTI. Starting with Figure 1, during the enrollment quarter (Q0) those in the “more successful” earnings group already average $1073 more in earnings per quarter than the “less successful” earnings group. This difference increased to $1710 by Q8. Most of the increased difference resulted from the upward movement of earnings for the “more successful” group by $630 (57%) relative to Q0. By contrast, the “less successful” groups average earnings level at Q8 was actually slightly lower ($8) at Q8 than at Q0. 31 Figure 1: Mean UI Earnings Q-4 through Q8 by UI Earnings Success Indicator (UI$I) UI Earnings per Quarter $2,000 $1,800 $1,600 UI Earnings in 1996 $ $1,400 $1,200 Less Successful More Successful $1,000 $800 $600 $400 $200 $0 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 Participation Quarter, 0 = Entry to WPTI 31 Though there was a small decline in average earnings for the “more successful” group over the final quarters of the data series, there was quite a large one, in percentage terms (55%), for the “less successful” group after they reached their maximum in Q3. The decline was small ($37) in absolute terms as average earnings for members of this group peaked at $67 per quarter. DRAFT – Do Not Cite Without Permission 12 Figure 2: Mean UI Employment Rate Q-4 through Q8 by UI Employment Success Indicator (UI$I) UI Employment Rate 100.0% 90.0% 80.0% Percent UI Employment 70.0% 60.0% Less Successful More Successful 50.0% 40.0% 30.0% 20.0% 10.0% 0.0% -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 Participation Quarter, 0 = Entry to WPTI Figure 2 displays the UI employment rate trends for the “more successful” and “less successful” groups for the UI Employment Success Indicator (UIEI). The general trends seen in Figure 1 are again present. There is a substantial difference in the Q0 employment rate of about 50 percentage points. This difference grows to 75 percentage points by Q8. Not unexpectedly, given the usually strong relationship between average earnings and employment, the quarter to quarter post enrollment trends for both groups are similar to those observed for earnings. The employment rate for the “less successful” is never very high, but the maximum (23%) is reached in Q4, followed by some decline. For the “more successful” group, the UI employment rate continues to rise through Q7, with some decline in the last quarter. We have not included a graph for the Form Employment Success Indicator (FEI) as those data exhibit the same basic pattern as that presented for the UIEI indicator in Figure 2. The main difference is the slightly higher employment rates observed in most quarters. Additionally, the information in both Figures 1 and 2 indicate that substantial divergence was present between the “more successful” and “less successful” groups for both the UI Earnings Success (UI$I) and UI Employment Success (UIEI) indicators in the quarters (-4 through -1) approaching that observed in the initial quarter of WPTI enrollment. 32 The trend lines for this period also show that the pattern of increasing divergence in earnings and employment outcomes was clearly present about a year before project entry. This is an important finding as it suggests to us that the WPTI 32 The pre-enrollment trend lines for the ”less successful” participants, especially for UI earnings, are consistent with the expectation of reduced work effort prior to entry to employment and training programs (the “Ashenfelter” dip). In marked contrast, the participants classified into the “more successful” groups are exhibiting strong trends toward increasing their work effort. Unfortunately we do not have attitudinal data for this period. Additionally, as will be reported later in this paper, we cannot distinguish the “more successful” from the “less successful” based on attitudinal data collected at WPTI enrollment, at least as structured in the data set we used. DRAFT – Do Not Cite Without Permission 13 intervention or any of its particular features is not likely to be the primary driver of the increasing difference between the “more successful” and “less successful” groups. Saying this, however, does not mean that WPTI did not help more successful participants to exploit whatever advantages they possessed when entering the project. Nonetheless, as there is no apparent increase in the rate of employment and earnings growth for those in the “more successful” groups in the quarters following enrollment, it is reasonable to ask why we think WPTI may have contributed to continued growth in the mean values observed in Figures 1 and 2. We offer three arguments. First, the employment rates and average earnings levels reach values far above what is normally observed for those participating in Social Security disability programs. While by itself this proves nothing, the finding suggests that something unusual has happened. Second, our models include variables that measure the impact of various features of the intervention. Should at least some of these variables have a strong and positive impact on outcomes, there would be reason to think the intervention contributed to the observed outcomes. This proved to be the case. Third, participants, through focus groups and follow-up surveys, suggested that the Pathways experience helped them in multiple ways to achieve better employment outcomes, including ongoing availability of useful information from a trusted source, continuing personal/social support, and skill development. A number of WPTI participants pointedly contrasted the usefulness of having an available source of assistance and support over the long haul, in contrast to their experience with other programs that had only offered a short-term fix (for example, job placement services). 33 We close this section by looking at whether those who were classified as “most successful” for one indicator would be so classified for the others. A quick look at Table 1 will confirm that the sets of the “most successful” are not perfectly equivalent. However, cross-tabulations between the indicators confirm substantial overlap, though least strongly between the UI earnings and forms employment indicators. 34 This is probably a result of the additional jobs identified through the update form (predominately self-employment) not being reported to the UI system. Spearman correlations also indicate substantial, but far from complete, overlap between the “most successful” groups. The correlation between the two UI based indicators was strongest at .691. The weakest association among the three pairs of variables was .459 between the UI$I and FEI indicators. Thus, with appropriate caution, we believe that it is reasonable to talk about participants as having, in a general way, “more successful” and “less successful” overall outcomes. Nonetheless, we will report findings only for the UI$I, UIEI, and FEI variables. 33 34 See Delin, Barry S., et al. 2004. pp.123-25. 72% of the “more successful” using the FEI indicator were in the “more successful” group for the UI$I indicator. 69% of those classified as “more successful” on the UI$I variable were also classified as among the “more successful” on the FEI variable. DRAFT – Do Not Cite Without Permission 14 Independent and Control Variables We utilized the same data set as was used for the WPTI outcomes evaluation. 35 This data base includes a wide range of administrative, encounter, and survey data. Although we had access to all of the data collected for the WPTI evaluation, given the exploratory purpose of this paper we limited this study to those variables included in the data set for the WPTI Final Project Report or, in several cases, to recoded or transformed versions of those variables. We planned to utilize independent and control variables in seven categories: (1) variables capturing some aspect of the WPTI service model or its provision, (2) variables capturing use of policy alterations created by or associated with WPTI, (3) sociodemographic variables, (4) benefit program participation variables, (5) work experience variables, (6) variables describing disability type and severity, and (7) participant attitudes and perceptions. We did not use all available variables. We began by selecting variables that had been found to have a significant impact on WPTI participant earnings and employment outcomes relative to those for the comparison group. We also, with one exception discussed in footnote #41, included variables that had a significant effect on outcomes among WPTI subgroups that had been examined for the WPTI Final Project Report. We then examined the bivariate relationships between these variables and the dependent variables with the intention of excluding variables when the appropriate measure(s) of association did not meet a significance level of .2. In some cases, we retained variables that did not meet this standard, either because they measured some aspect of the intervention we thought critical or, as in the case of certain socio-demographic variables, we thought exclusion would have been questioned. What follows is a listing, by area, of the seventeen independent and control variables used in our regression models. In a few cases we include some discussion of these choices or of those not made. Fuller descriptions of these variables and of associated frequencies and/or basic statistics can be found in Appendices A through C. WPTI service provision: 36 Hours of benefits counseling provided (variable name = BCtoQ5) Hours of vocational services provided (VStoQ5) 35 Data elements included both those required by the SPI Project Office at Virginia Commonwealth University and others specific to the WPTI evaluation and/or that of the closely related 3-State Work Incentives Initiative. Pamela Hanes had the lead role in designing both the WPTI and 3-State evaluations. Among the many individuals with some role in putting together the WPTI data set, we wish to give special recognition to Janne Boone and Julia Drew. We also wish to thank David Sage for his contributions to both the WPTI evaluation and this study. 36 The WPTI data set we used includes hours of service delivery in two formats: Q0 through Q2 and Q0 through Q5. We chose to use the longer duration because of WPTI’s emphasis on getting follow-up services at need and the fact that many participants had to wait before receiving services (especially benefits counseling) because of staff attrition or DVR budgetary problems. When we looked at the strength and significance of measures of association between the service provision and dependent variables, the bivariate relationships tended to be stronger with the Q0 through Q5 versions of the service provision variables. DRAFT – Do Not Cite Without Permission 15 Hours of job coaching services provided (JCtoQ5) Implementation Quality (IMPQUAL) WPTI policy alterations: SSI Waiver (SSIWAV) Medicaid Buy-in (MAPP) Socio-demographic: Age (AGE) Sex (SEX) Race (RACE2) Education (EDUCrev) Benefit program participation: 37 SSDI/SSI status (SSACat) 38 37 With the exception of the SSDI/SSI status variable, we view the program participation variables as control variables. The SSA Benefit Amount and Food Stamps participation, both at Q0, are intended to control for the potential individual level tradeoffs each new WPTI participant faced as a consequence of entering a project that had as its purpose increasing that individual’s work effort. For example, consider the different situations (all else being equal) faced by a SSI only recipient and a SSDI only beneficiary receiving a benefits check well over SGA. The SSI recipient’s check will be significantly less than SGA and will only be subject to a $1 reduction for each $2 earned, irrespective of whether or not earnings rise above SGA. By contrast for the SSDI beneficiary any earnings over SGA, after completion of the Trial Work Period, result in the full loss of the benefit check. To at least break even, the beneficiary must earn her/his former benefit amount plus the SGA amount. In the real world, these calculations are further complicated by loss of “leisure” time (no trivial consideration for many with a severe disability) and potential loss of other public benefits. More WPTI participants utilized Food Stamps than any other non-health care related public program; participation at Q0 is used as a proxy for whether the participant faced potential loss of non-Social Security benefits from increased work activity. The third of these variables, income change between the base and outcome year, is intended to control for the actual trade off between earnings and the Social Security benefit amount. The SSA Benefit Amount at Q0 and income change between the base and outcome years are monetary amounts. These variables were standardized because of their large scale relative to the values of other variables used in the regression models. Additionally, the underlying measure of income change is a proxy rather than an actual value. It includes UI earnings and the income received from Social Security and the Wisconsin State SSI Supplement and thus excludes other possible sources of income. Moreover, the proxy excludes income received by any other member of a participant’s household. 38 An individual may participate in both the SSI and SSDI programs at the same time, most often because that individual’s SSDI benefit is less than the SSI FBR amount. We chose to use a variable that classified SSDI only and concurrent SSI/SSDI cases in the same category, all SSI only cases in a second. This reflects our judgment that the overall impact of SSDI policies on earnings is stronger than that of SSI policies. When a concurrent beneficiary works, the size of their benefit check reflects the application of both SSDI and SSI rules. In most cases, the impact of the SSDI rules is more powerful. In particular, concurrent beneficiaries face the same constraints on earnings above SGA following completion of the Trial Work Period as a SSDI only beneficiary. Any earnings over SGA result in a 100% loss of the SSDI portion of a concurrent beneficiary’s Social Security check. DRAFT – Do Not Cite Without Permission 16 Food Stamps (FSQ0) SSA Benefit Amount at Q0 (Q0BS$_sdz) Income change between the base and outcome years (Chg_sdz) 39 Work experience: Post Disability Employment Ratio (EMPRAT) 40 Disability type and severity: Primary Disability (PRIMDIS) 41 DVR Order of Selection Category (OOSrev) 42 To our disappointment, we did not include any measures of participant attitudes or perceptions at program entry in our modeling. We had expected these variables to explain at least some of the difference in employment outcomes, especially the differences in earnings and employment rates that were exhibited even before the quarter of project entry. To our surprise, we saw almost no evidence that participants classified into a “more successful” group (for any of the three indicator variables) differed from participants classified into a “less successful” group on any of the 39 In the WPTI final report “income” was treated as a dependent variable. In this study, we treat it as a control variable, reasoning that the severity of the tradeoff between earnings and cash benefits might affect participant decisions about work. Policy features of the WPTI intervention such as the SSI Waiver and MAPP were intended to decrease the severity of the tradeoff. Benefits counseling was intended to help participants to make informed choices, including whether and to what degree to use these policy features and other available work incentives. 40 “Post-disability,” in this context, begins at the eligibility date for Social Security benefits, not the actual date that the impairing condition started. The EMPRAT variable is a proxy for the actual ratio of time employed between entering a Social Security benefit program and entry into WPTI. 41 Assignment was to one of the three disability groups was based on the participant’s RSA-911 code in their DVR record. Assignment was according to a crosswalk developed by Pamela Hanes and her staff at Oregon Health and Sciences University and modified by the Wisconsin based WPTI evaluation staff. There was very considerable overlap between these disability categories and the type of WPTI agency at which a participant enrolled. This was largely a product of the secondary WPTI eligibility criteria that allowed each of the four agency types (Physical Disability, Mental Health, Developmental Disabilities, and HIV/AIDS) to enroll participants who fit within one of these categories as defined by their contracts with DHFS. As both variables more or less measure the same thing, we included one. We chose to use the one that was the more direct indicator of a disabling condition as we are interested in the impact of attributes that participants brought with them to WPTI. However, this choice has a cost. There were very large differences in the average employment rates and earnings across the four agency types, differences that on the basis of interviews and focus groups conducted during WPTI appear to be related to agencies’ missions and service philosophies. See Delin, Barry S., et al. 2004. pp. 134-40 for differences in outcomes by WPTI agency type, pp. 17-23 for differences in agency service philosophy by WPTI agency type. 42 The DVR OOS code is used as a proxy for severity, though one that reflects that agency’s concern with assessing how difficult it will be to prepare an individual to reenter (or initially enter) the labor force. DRAFT – Do Not Cite Without Permission 17 attitudinal/perceptual measures, at least at the time of entry to WPTI. Our data set included measures of work motivation, perceived barriers to work, self-esteem, mastery, and subjective physical health (PCS12) and mental health (MCS12) scales from the SF36 assessment instrument. Only one combination of the attitudinal and dependent variables had an ordinal by ordinal measure of association lower than .3; most values were .7 or higher. The single exception was the association measure between UI$IP and the MCS12 score which just met the .2 significance level we specified for including variables in our regression models. However, we did not include this variable in the modeling because there were sixty-three missing cases. 43 Analysis Methods This study uses binary logistic regression to identify the variables which seem to most strongly account for the distribution of WPTI participants into the “more successful” and “less successful” categories for each of the three dependent variables (UI$I, UIEI), and FEI). The analysis seeks to account for outcomes over either a thirty month (for the UI data based indicators) or a twenty-seven month period (for FEI), though the analysis is not set up as a time series. Only a maximum of 483 of the 506 cases are available due to missing data for two of the seventeen independent variables. 44 Given the exploratory purpose of this study, we decided to use forward selection models utilizing the 483 cases. We also ran models after removal of outliers at or above 2.58 standardized residual level. When there are no major differences between the original models and those with the outliers removed, we will present the original models. 45 The standard .05 significance level is criterion used for rejecting null hypotheses. 43 These 63 cases represent about 12% of the 506 cases. Indeed, the use of any of the attitudinal variables would have reduced our N by a minimum of 38 cases. Nonetheless, we would have used at least some of these variables in our models if there had been any indication that they might have a meaningful effect. 44 Given the exploratory nature of this study, we decided to keep the modeling as simple as possible. As noted earlier we were willing to ignore quarter to quarter variation in this initial analysis, because it looked at a relatively short period for a population unusually subject, especially because of variation in health and the availability of support services, to variability in their work effort. We were also concerned about the technical difficulties of modeling when one of the two analytical groups had majorities with no employment or earnings in every calendar quarter. Two independent variables had missing cases. There were 21 missing values for OOSrev and 5 for EDUCrev. As 3 cases had missing information for both these variables, the actual number of excluded cases was 23, 4.5% of the original 506. 45 Models were run using SPSS for Windows V. 13 using the LR option to calculate the models. We used the default values offered by SPSS, including those for inclusion (.05) and exclusion (.10) of variables at each step. We also ran models using backward selection to check for consistency between inclusion of variables and their relative strength between the forward and backward selection models. While there were two cases among five general models run where the backward selection models included variables not chosen in forward selection, these variables were weak and either of borderline significance or non-significant. DRAFT – Do Not Cite Without Permission 18 In modeling the UI Earnings Success Indicator (UI$I) we used all seventeen independent variables. For the two employment indictors, (UIEI and FEI) we did not include the variable measuring hours of job coaching in the models because that variable (JCtoQ5) measures service provision that, in principle, should have occurred only after a participant became employed. We faced a more difficult choice about whether to include the Medicaid Buy-in participation variable (MAPP) in the employment models. As a Balanced Budget Act authorized Buy-in, the Wisconsin program (MAPP) requires employment as a condition of participation. The one exception to this requirement is participation on a time limited basis in an approved employment preparation program (HEC). At any time, only a handful of MAPP participants are using HEC. As such, it would seem reasonable to exclude MAPP from the employment models. However, MAPP does not require any monetary earnings or minimum hours of work as a condition of either initial or continuing eligibility. Employment that provides inkind compensation is acceptable. So, apparently, is any number of work hours above zero, provided some work is performed every month. Though such individuals are eligible for the Wisconsin Buy-in, they do not meet the definition of employment used in WPTI or the other SPI projects. 46 Moreover, the fact that a significant proportion of MAPP participants report either no earnings (12%) or no more than $67 per month (50%) suggests that there is a difference between MAPP participation in itself and the attainment of the relatively high employment outcomes that WPTI and, ultimately, the Wisconsin legislature and U.S. Congress, hoped to motivate. 47 Given the ambiguity concerning whether MAPP should be included in or excluded from the employment models, we decided to run employment models both ways. To foreshadow our findings, MAPP participation proves to be an extremely strong factor in determining whether participants are in the “most successful” group in all models that include the MAPP variable. Finally, our models include three categorical variables, primary disability (PRIMDIS), educational attainment (EDUCrev) and WPTI implementation quality (IMPQUAL) that have more than two values. The reference category is “physical disability” for PRIMDIS, “more than high school” for EDUCrev, and “highest” for IMPQUAL. 46 Employment in all SPI projects, including WPTI, was defined as requiring monetary compensation and being continuous, rather than episodic, in nature. The SPI definition did not identify minima, but did include some examples of situations (e.g., receiving payment for occasional odd jobs outside of contracting such work on a continuing basis) that did not meet the definition. 47 APS Healthcare, Inc. 2005. Medicaid Purchase Plan Evaluation Annual Report (for 2004). Madison WI: APS Healthcare, Inc. pp. 11-12. The values cited are based on information collected by county economic support workers for October 2004. DRAFT – Do Not Cite Without Permission 19 Logistic Model Results Tables 2 through 6 provide results for the regression models we tested that are the “best” representations of the relationships between our indicator variables. We have placed “best” in quotation marks to signal that as a group these models, especially those using the encounter data (FEI), are not particularly strong as explanations of what distinguishes the “more successful” from the “less successful” for each of our dichotomous outcomes. Nonetheless, we argue that the results can provide some useful information to policymakers and stakeholders as to whom might be targeted for participation in return to work programs and, at least in the area of policy modifications, what might support the employment efforts of persons with disabilities who participate in the Social Security income support programs. Our discussion of Tables 2 through 6 focuses on the final two columns in each table. The significance (Sig.) value indicates whether a variable has been correctly included in the model and that the effect size reported for any independent variable is credible. As previously noted, our criterion is a value no larger than .05; our discussion, therefore, concentrates on variables that meet this standard. Exp(B) represents the “odds ratio.” When the dependent variables are dichotomous, as they are in this study, an Exp(B) of “X” can be understood as a change in the odds any case will have one value of the dependent variable rather than the other. For example, in Table 2, the Exp(B) value of 3.486 indicates, all else held constant, that a person signing up for the SSI waiver had about 3.5 times the odds of being in the “more successful” earnings group than in the “less successful” one. 48 We will use the Exp(B) to communicate the direction and relative strength of a variable relative to others in a model. Table 2 presents information about the UI Earnings Success Indicator (UI$I) model. The UI$I model is statistically significant and, according to the Hosmer and Lemeshow test, exhibits adequate “goodness of fit.” 49 The model appears to explain a usefully large amount of variance (Nagelkerke R Square = .341). The data in Table 2 indicate that WPTI’s policy features have strong and positive effects, resulting in odds of about 3.5 to1 that participants using either the SSI waiver or MAPP had earnings outcomes that placed them in the “more successful” category. By contrast, most WPTI service variables are excluded from the model. The one exception, the hours of job coaching provided, appears to have a very modest positive effect. 48 When the dependent variable has two values, an Exp(B) of 1 indicates no difference in the odds of being in one group rather than the other. Exp(B) values greater than 1 indicate a positive relationship between the independent and dependent variables, values less than 1 an inverse relationship. The further the Exp(B) value is (as a ratio) in either direction from 1, the greater the independent variable’s impact. Additionally, when a categorical variable has more than two values, the Exp(B)s are interpreted in relation to a reference category. 49 However, the classification table for this and our other models do not indicate that the models are especially good at correctly predicting the assignment of cases to their actual dependent variable category. The correct classification rate for the UI$I model is typical with a value of about 77%. DRAFT – Do Not Cite Without Permission 20 Table 2: Model Results for UI Earnings Success Indicator (UI$I). N=483. Variable B S.E. Wald Sig. Exp(B) PRIMDIS -.134 .304 .195 .659 .874 (cognitive/DD) PRIMDIS .729 .271 7.264 .007 2.073 (affective/MH) OOSrev .477 .216 4.852 .028 1.611 SSACat -969 .282 11.812 .001 .380 SSIWAV 1.249 .257 23.615 .000 3.486 MAPP 1.260 .307 16.872 .000 3.524 EMPRAT .670 .231 8.410 .004 1.954 JCtoQ5 .035 .015 5.421 .020 1.036 Q0BS$_sdz -.302 .126 5.733 .017 .739 FSQ0 -.715 .271 6.953 .008 .489 Chg_sdz .532 .127 17.524 .000 1.701 Constant -6.03 .514 1.374 .241 .547 Variables not included by forward selection: AGE, SEX, RACE2, EDUCrev, IMPQUAL, BCtoQ5, and VStoQ5. Other variables that appear to have a strong positive effect on motivating inclusion in the “more successful” earnings group include working a larger proportion of time since becoming qualified for a Social Security program (EMPRAT, Exp(B) = 1.954), having a less severe disability (OOSrev, , Exp(B) = 1.611), and having a primary disability (PRIMDIS) classified as “affective/mental” (Exp(B) = 2.073) The first two findings are consistent with expectations, the positive result for “affective/mental” conditions compared to the reference category of physical disabilities is perhaps more controversial. At least in the past, those with a mental illness (as a primary disability) were somewhat more likely than those with a physical disability to participate in SSI rather than SSDI. Those in SSI were thought to have less earnings capacity because of their more limited work experience. Indeed, the Exp(B) value for SSACat (.380) indicates a strong, but inverse relationship between being a SSI recipient and the odds of being included in the “more successful group.” All three of the benefits related control variables are included in the model and have Exp(B) values consistent with our suppositions. Not unexpectedly, increases in the income change variable (Chg_sdz, Exp(B) = 1.701) motivate inclusion into the “more successful” group. This is good news as it suggests that earnings gains more than offset losses in income from Social Security and the SSI State Supplement. The Exp(B) values well below 1 for the Social Security/state supplement amount (Q0BS$_sdz, Exp(B) = .739) and food stamps participation (FSQ0, Exp(B) = .489) are in concert with our expectation that those with either higher levels of income support or relying on non-SSA benefits programs would be less likely to achieve earnings outcomes that would place them in the “more successful” category. Finally, we were surprised that none of the socio-demographic variables were selected in the UI$I model. Most notable is the absence of educational attainment DRAFT – Do Not Cite Without Permission 21 (EDUCrev) given the strong and apparently strengthening relationship between educational attainment and earnings. 50 Tables 3 and 4 present results for the UI Employment Success Indicator (UIEI) models from which outliers were removed. For reasons discussed in the previous section of this paper, we are presenting models both including and excluding Medicaid Buy-in (MAPP) participation. Table 3: Model Results for UI Employment Success Indicator (UIEI) with MAPP. Outliers Removed, N=472. Variable B S.E. Wald Sig. Exp(B) PRIMDIS 1.064 .309 11.842 .001 2.899 (cognitive/DD) PRIMDIS 1.086 .284 14.588 .000 2.963 (affective/MH) SSACat -.694 .290 5.719 .017 .499 SSIWAV 1.228 .275 19.892 .000 3.416 MAPP 2.006 .327 37.582 .000 7.436 EMPRAT .710 .261 7.410 .006 2.003 Chg_sdz 1.215 .176 47.792 .000 3.370 Constant -1.606 .430 13.943 .000 .201 Variables not included by forward selection: AGE, SEX, RACE2, EDUCrev, OOSrev, IMPQUAL, BCtoQ5, VStoQ5, Q0BS$_sdz, and FSQ0. The results for the model including MAPP are found in Table 3. This model exhibited adequate goodness of fit and had a Nagelkerke R Square of .391. Though fewer variables were selected than for the UI employment model, many of the same independent variables exhibit strong and positive relationships with the dependent variable. The policy components of the WPTI intervention have, if anything, an even stronger effect than they have in the UI earnings model. The Exp(B) value for SSI Waiver use (SSIWAV) is over 3.4, for MAPP over 7.4. Intervention service variables are again conspicuous by their absence, as are the socio-demographic variables. Once again, having a strong employment history after entry into SSDI or SSI (EMPRAT) strongly increases the odds a participant is in the “more successful” category (Exp(B) = 2.003). Participants who were SSI only again have reduced odds of being in the “more successful” category. This negative impact (Exp(B) = .499) is not quite as strong as for the UI earnings model (Exp(B) = .380). Two of the benefits related control variables are absent from this model, but the remaining one, the income change variable (Chg_sdz), has about twice the impact on the odds of being in the “more successful” group than it did in the UI earnings model. 50 See Hanushek, Eric A. 1996. “Outcomes, Costs, and Incentives in Schools” in eds. Hanushek, Eric A. and Jorgenson, Dale W. Improving America’s Schools: The Role of Incentives. Washington DC: National Academy Press. pp. 31-33. DRAFT – Do Not Cite Without Permission 22 Perhaps the most salient difference between the UI Earnings Success model and this employment model (as well as the other employment models) is that inclusion in the “cognitive/developmental” category of PRIMDIS strongly increases the odds of being in the “more successful” category. The Exp(B) value is about 2.9. This finding was not expected for much the same reasons discussed in regard to the surprising strength of the “affective/mental” category in our discussion of Table 2 findings. Indeed, our expectation is that those with cognitive/developmental disabilities would tend to have even less work experience and education in comparison to those with physical disabilities. 51 Finally, inclusion in the “affective/mental” category again strongly increases the odds of inclusion among those exhibiting more successful outcomes. Turning to Table 4, which shows results for the UIEI model run without MAPP, we find a pattern of results generally similar to those observed for the UIEI model with MAPP. This model had a slightly lower Nagelkerke R Square of .343. Table 4: Model Results for UI Employment Success Indicator (UIEI) without MAPP. Outliers Removed, N=472. Variable B S.E. Wald Sig. Exp(B) PRIMDIS .653 .304 4.611 .032 1.922 (cognitive/DD) PRIMDIS .975 .281 12.082 .001 2.652 (affective/MH) RACE2 .972 .321 9.176 .002 2.643 SSACat -.944 .307 9.491 .002 .389 SSIWAV .817 .264 9.596 .002 2.264 EMPRAT .803 .254 9.970 .002 2.232 BCtoQ5 .008 .004 4.802 .028 1.008 Q0BS$_sdz -.333 .143 5.418 .020 .717 Chg_sdz 1.048 .157 44.635 .000 2.853 Constant -2.759 .782 12.445 .000 .063 Variables not included by forward selection: AGE, SEX, EDUCrev, OOSrev, IMPQUAL, BCtoQ5, VStoQ5, and FSQ0. The most notable addition to this model is the dichotomous race variable (RACE2), the first time that a socio-demographic variable appears in one of the models. The Exp(B) value indicates that participants who identified themselves as “white” had about 2.5 the odds of being in the “more successful” UIEI employment group than someone who had self-identified as having a different racial background. 52 A WPTI intervention service component (benefits counseling hours) also enters the model, but has essentially no effect on outcomes. The post-disability work history variable (EMPRAT), both primary disability categories (PRIMDIS = affective/MH or cognitive/DD), SSA Waiver use (SSIWAV), and the change of income variable (Chg_sdz) all remain strongly associated with having a 51 52 See Table 10 for information about educational attainment. More than three-fifths of the “other” group identified themselves as “black.” The next largest group was composed of those who saw themselves as having multiple racial identities. DRAFT – Do Not Cite Without Permission 23 positive outcome. Though the effect of the “cognitive/developmental” category of PRIMDIS is noticeably weaker than in the UI employment (UIEI) model that included MAPP, the effect of being in the SSI waiver is stronger. SSI only participants again have greater odds of poor outcomes than SSDI or concurrent beneficiaries; the strength of the effect appears somewhat stronger than in the other UIEI model. Finally, the SSA benefits level at enrollment variable (Q0BS$_sdz) is included in this model. As expected, Q0BS$_sdz (Exp(B) = .717) is inversely related to inclusion in the “more successful” group; the variable has about the same impact it did in the UI Earnings Success model. Tables 5 and 6 present results for the Employment Success Indicator based on information from the WPTI Enrollment and Monthly Update forms. These models, while exhibiting adequate “goodness of fit,” explained much less of the variance than the UI$IP and UIEI models presented in Tables 2 through 4. The Nagelkerke R Square for the model including MAPP is .197, without MAPP .127. We first present findings from the FEI model that included MAPP. Table 5: Model Results for “Form” Employment Success Indicator (FEI) with MAPP. N=483. Variable B S.E. Wald Sig. Exp(B) PRIMDIS .833 .266 9.776 .002 2.299 (cognitive/DD) PRIMDIS .550 .253 4.732 .030 1.733 (affective/MH) SSIWAV .628 .225 7.803 .005 1.875 MAPP 1.442 .293 24.244 .000 4.228 EMPRAT .707 .219 10.465 .001 2.028 IMPQUAL -.370 .305 1.475 .225 .690 (low) IMPQUAL -676 .221 9.343 .002 .509 (medium) Chg_sdz .242 .108 5.052 .025 1.274 Constant -.978 .216 20.415 .000 .376 Variables not included by forward selection: AGE, SEX, RACE2, EDUCrev, OOSrev, SSAcat, BCtoQ5, VStoQ5, Q0BS$_sdz, and FSQ0. As with the UI Employment Success Indicator model that includes MAPP, both Buy-in and SSI Waiver participation are strong (Exp(B) = 4.228 and 1.875 respectively) and thus lead to substantially increased odds of inclusion in the FEI “more successful” category, albeit that the Exp(B) values are considerably lower than the ones for the comparable UIEI model. Much the same can be said for the two primary disability (PRIMDIS) categories and the income change control, Chg_sdz. However EMPRAT, the work history variable, is about equally strong and positively associated with desired outcomes in both the employment models that include the MAPP variable. The FEI model (including MAPP) does not include the Social Security program variable (SSACat), but for the first (and only) time the implementation quality variable (IMPQUAL) appears in a model. While the coefficients are in the anticipated direction (that is, participants at agencies rated as having “low” or medium” implementation quality DRAFT – Do Not Cite Without Permission 24 have lower odds of being in the “more successful” category than those at agencies rated “high”), only one of the coefficients is statistically significant. The FEI model that excluded the MAPP variable (see Table 6) presents the most idiosyncratic results, probably reflecting its slight explanatory power. It is the only model that excludes SSI Waiver use (SSIWAV). It also excludes the race variable that was quite strongly related to good employment outcomes in the comparable UI Employment Success Indicator model. Table 6: Model Results for “Form” Employment Success Indicator (FEI) without MAPP. N=483. Variable B S.E. Wald Sig. Exp(B) PRIMDIS .675 .256 6.928 .008 1.963 (cognitive/DD) PRIMDIS .608 .240 6.412 .011 1.837 (affective/MH) EMPRAT .877 .209 17.541 .000 2.403 BCtoQ5 .010 .003 8.014 .005 1.010 Chg_sdz .311 .100 9.560 .002 1.364 Constant -1.286 .206 39.157 .000 .276 Variables not included by forward selection: AGE, SEX, RACE2, EDUCrev, OOSrev, SSAcat, IMPQUAL, VStoQ5, SSIWAV, Q0BS$_sdz, and FSQ0. More typically, both of the primary disability categories, work history (EMPRAT) and income change (Chg_sdz), are present and positively related to inclusion in the “more successful group” at strengths similar to the other FEI model. Also, like the other FEI model, the Social Security program variable (SSACat) is not selected into the model. Finally, benefits counseling hours is included and significant, but the Exp(B) value is so close to “1” that the variable appears to have but a trivial impact on results. Discussion In this section we review the impact of our independent and control variables on inclusion into the more and less successful categories of our three dichotomous dependent variables. Using descriptive data presented in Appendices B and C and some material from the WPTI Final Project Report, we identify and discuss possible explanations for unexpected results from our regression models. We also make some suggestions for further analysis of the WPTI dataset. First, however, we briefly examine whether classification into one of the “more successful” categories of the earnings and employment indicators results in the participant being “better off” because of their experience in WPTI. We define “better off” as having a higher income, not just higher earnings or a higher probability of employment. Not only is this a pertinent standard for WPTI’s value from a participant standpoint, it also speaks to the issue of whether the gains in earnings and the reductions in benefit costs desired by both Wisconsin and SSA could be expected to be lasting should the “more successful” participants retain the capacity and opportunity to maintain increased work effort. The data in Tables 7, 8, and 9 suggest that the desired patterns of income growth, including earnings growth and benefit cost reduction are DRAFT – Do Not Cite Without Permission 25 present for those in the “more successful” categories. As previously identified (see footnote #37), our income measure is a proxy that is limited to the participant’s UI earnings and Social Security and State SSI Supplement cash benefits. Table 7: Change in Income (Earnings and Social Security/State Supplement Amount) between the Base and Outcome Years (1996 GDP$) Variable Mean Change Median Change Standard Dev. “More Successful” UI Earnings $1525 $1033 $5925 UI Employment $2251 $1100 $5963 Form Employment $1230 $384 $5031 “Less Successful” UI Earnings -$327 $50 $2041 UI Employment -$361 $59 $2945 Form Employment -$22 $70 $3743 Table 8: Change in UI Earnings between the Base and Outcome Years (1996 GDP$) Variable Mean Change Median Change Standard Dev. “More Successful” UI Earnings $2134 $1423 $6702 UI Employment $2899 $1343 $6474 Form Employment $1823 $519 $5524 “Less Successful” UI Earnings -$86 $0 $457 UI Employment -$72 $0 $2998 Form Employment $244 $0 $3812 Table 9: Change in Social Security/State Supplement Amount between the Base and Outcome Years (1996 GDP$) Variable Mean Change Median Change Standard Dev. “More Successful” UI Earnings -$609 $51 $2936 UI Employment -$647 $0 $2786 Form Employment -$594 $39 $2468 “Less Successful” UI Earnings -$241 $70 $1959 UI Employment -$289 $76 $2278 Form Employment -$266 $78 $2462 Another prominent result is that for every mean presented in these tables the standard deviation is much larger than the mean. Moreover, the standard deviations for income and earnings change for each of the “more successful” groupings are very high on an absolute basis. These results signify that the dichotomous variables obscure a large amount of variation in UI earnings among both the “more successful” and “less successful,” but especially among the former. Though we explained our reasons for using dichotomous indicators in this exploratory effort and presented evidence showing that the “more successful” and “less successful” groups were easily distinguishable, it DRAFT – Do Not Cite Without Permission 26 would make sense to investigate the use of indicator variables with more than two levels and the use of multinomial logistic methods. 53 Returning to the WPTI intervention, we begin by noting that differences between the more and less successful on all three of our indicator variables were apparent well before enrollment in WPTI. More importantly, the growing divergence between the more and less successful groups on each indicator grew at about the same rates in the quarters before and after enrollment, with increasing outcomes among the “more successful” responsible for almost all of the growing difference (see Figures 1 and 2). Therefore, we think it more appropriate to think about WPTI as a program that supported continued growth in outcomes rather than as the principal cause of those outcomes. 54 None of the WPTI service provision variables, including the one measure of (perceived) intervention quality, appear to have directly contributed to whether participants ended up in the more or less successful categories at a level that we would consider meaningful. In brief, we hypothesize that any effect of the delivered quality of WPTI services (IMPQUAL) was small compared to differences rooted in WPTI agency recruitment patterns and service delivery philosophies. We will discuss this in more detail when we look at the primary disability (PRIMDIS) variable. There may also be PRIMDIS specific patterns in the value of different amounts of provision of specific services for promoting better outcomes (e.g. job coaching for those in the “cognitive/developmental” category within PRIMDIS), but at the WPTI program level any impact of these variables is lost in the relatively high variation in service hours provided to both “more successful” and “less successful” participants. 55 By contrast, the impacts of WPTI’s policy features, the SSI waiver and MAPP, were generally strong. In most models, use of one of these policy features made the 53 We were concerned that our control variable for income change, Chg_sdz, had such a large earnings component that it weakened the explanatory power of the models. Despite feeling it was conceptually less appropriate than “income change,” we tried a standardized change in benefits (Social Security and state supplement) amount variable in our models. In no case did doing so improve the explanatory power of the model or substantially change the inclusion/exclusion of other variables. 54 It is apparent from interviews and focus group that some agency staff and early participants viewed WPTI more as a test of policy options that would ameliorate the tradeoffs between employment and access to public benefits than as an employment program. In addition to such reports, this assertion gains support from the comparatively high employment rates of the earliest WPTI enrollees compared to those who entered somewhat later. Participants who enrolled in the first 9 months of the project had about a 15% higher employment rate at entry than those who enrolled in the next nine months, a fact that cannot be accounted for by the modest decline in labor market conditions that occurred over this time period (July 1999 through December 2000). See Delin, Barry S., et al. 2004. pp. 193-94 and 29-30. 55 The WPTI evaluation was not set up to clearly distinguish the impact of key service components, especially benefits counseling, from that of the intervention as a whole. However, using data from Vermont’s SPI program, one study found that the delivery of benefits counseling services was strongly and positively associated with subsequent employment and earnings outcomes. See Tremblay, Tim, Xie, Haiyi, Smith, James, and Drake, Robert. 2004. “The Impact of Specialized Benefits Counseling Services on Social Security Administration Disability Beneficiaries in Vermont,” Journal of Rehabilitation, 70, No. 2, pp. 5-11. DRAFT – Do Not Cite Without Permission 27 odds of being in a “more successful” group two to four times greater. This suggests that well crafted changes to the structure of incentives and disincentives built into public benefits programs (whether in a single program or, preferably in theory, across multiple ones) may prove a powerful intervention approach. However, given what we know about the modest employment outcomes associated with MAPP participation in general, we speculate that some of the services and supports provided through WPTI (e.g. benefits counseling) may have had an important role in generating the strong outcomes observed for WPTI participants who were also MAPP participants. 56 This claim receives support from participant survey and focus group data. We learned that participants vary greatly in when they felt ready to make use of the SSI Waiver, MAPP, or indeed other available work incentives. Readiness could vary for many reasons, including health and personal or familial circumstances, not directly related to employment goals or work activity. 57 This interpretation also receives support from our interviews with both community agency staff and DHFS and DVR administrators in a position to observe WPTI operations. As mentioned earlier, we were surprised at the lack of impact that sociodemographic variables had in our models, most particularly educational attainment (EDUCRev). Indeed, the tables in Appendix B show only small differences in the proportions in each category of EDUCrev between the “more successful” and “less successful” groups for any of the dependent variables. We thought it possible that there might be an unusually large number of very severely disabled participants with high educational attainment that would obscure the expected relationship between education and strong employment outcomes. The data in Table 10 does not support this possibility, but it does suggest another. Table 10: Educational Attainment (EDUCrev) by DVR Order of Selection (OOSRev) and Primary Disability (PRIMDIS) Variable Less than High High School or More than High School GED School OOSrev Most significant 54 (21%) 91 (36%) 108 (43%) Other 37 (16%) 54 (24%) 139 (60%) PRIMDIS Cognitive/developmental 19 (20%) 43 (46%) 31 (33%) Affective/mental 25(23%) 35 (32%) 49 (45%) Physical 92 (18%) 151(30%) 258 (52%) Persons classified as having a “physical” primary disability had a somewhat higher percentage in the “more than high school” category of EDUCrev than those classified as having an “affective/mental” condition as their primary disability. Table 10 also shows that WPTI participants with a physical condition as their primary disability had, as a group, much higher levels of educational attainment than those with a 56 However, caution is advised in interpreting the strong beneficial results associated with MAPP participation as only 16% of WPTI participants were enrolled in the Buy-in program. Even in the “more successful” groups the participation rates varied between 24% and 30%. 57 We have not yet looked at the participant follow-up survey data based on the dichotomous outcome variables we created for this study. Thus the characterization of results provided on page 13 reflects the dominant response pattern among those participants who returned surveys. The focus group data is much richer, but the number of cases is very small. DRAFT – Do Not Cite Without Permission 28 classification of “cognitive/developmental.” These distributions are not unexpected. What is unexpected is that, as a look at the PRIMDIS distributions in the Appendix B tables will indicate, the proportion of those with “physical” primary disabilities in each of the “more successful” distributions is always less than for the corresponding “less successful” distributions. By contrast, those with “affective/mental” primary disabilities show a very strong pattern in the opposite direction, as do, to a much lesser degree, those with cognitive/developmental primary disabilities. As with the seemingly anomalous results for the WPTI implementation quality variable, we will discuss this issue further in the context of the PRIMDIS variable. It is also possible that we have used a poor categorization for measuring educational attainment in the context of the contemporary labor market. As a high school diploma or GED increasingly becomes the effective minimum credential for work that pays a “living wage” or even work which opens up that possibility, we might have been better served by having a variable that distinguishes between different amounts of postsecondary education. Fortunately, the data to do this are available, though not in the data set used for this paper. We have little to add to our previous comments about our benefit program participation and work experience variables. As a group, their impact was much as we anticipated. There was an inverse relationship between receiving high levels of SSA and/or other income support benefits and those on SSI only were less likely to achieve earnings and employment outcomes that would result in assignment to the “more successful” groups. Also, it was not surprising to learn that those employed in a greater proportion of the years between entering Social Security benefit status and entering WPTI (EMPRAT) were more likely to have strong employment outcomes. The Exp(B) values for EMPRAT varied over a fairly small range (about 1.9 to 2.4) across all of the five logistic models. However, we observed that the distribution of EMPRAT was strongly bimodal, with more than 80% of the 506 cases either having reported earnings in none or all of the years between entering a Social Security disability program and entering WPTI. The bimodality was also observable in each of the “more successful” and “less successful” groups. Consistent with expectations, the “always employed” peak (EMPRAT = 1) was in all cases substantially higher than the “never employed” peak (EMPRAT = 0) for the “more successful,” and visa versa for the “less successful.” We do not know whether the distribution of EMPRAT across the Social Security (adult) disability program population would be similar; but if it is, it suggests a potential targeting or recruitment approach for return to work programs. In all of the models, those with an “affective/mental” primary disability showed a strong tendency to be included in the “more successful” groups compared to those in the reference PRIMDIS category, “physical.” This was also true for those in the “cognitive/developmental” category for all of the employment models, but not the earnings model. Though assignment of cases to a PRIMDIS category was based on information (RSA-911 codes) external to and usually predating WPTI enrollment, we urge caution in interpreting the relatively high effect sizes (Exp(B) from about 1.7 to 3.0) as an indicator of “physical” disabilities being a direct cause of less favorable employment outcomes. WPTI participants with a “physical” primary disability had higher level of educational attainment than either of the other disability categories. On other variables of interest such as SSA status, level of severity (OOSrev), and employment history since entering a SSA disability program, the distributions for the “physical” category of PRIMDIS were similar to those for persons in the “affective/mental” category DRAFT – Do Not Cite Without Permission 29 and much more promising than for persons in the “cognitive/developmental” category. Finally, we know that persons enrolled at WPTI “physical disability” agencies (most of whom were classified as having a “physical” primary disability) had, when employed, much higher average earnings than employed participants enrolled at other types of WPTI agencies. 58 There is nothing here to suggest why those with a PRIMDIS of “physical” should be so much less likely to be included among the “more successful.” Nonetheless, this result could have been anticipated from the results of the WPTI evaluation. In the regression models that examined differences among the three PRIMDIS disability groups within WPTI, those in the “affective/mental” and the “cognitive/developmental” categories had significantly better employment and earnings outcomes than those in the “physical” category.” The overall trends for the groups were broadly similar and the degree of divergence in estimated outcomes tended to persist over the full two year period following WPTI enrollment. 59 Previously we noted (see footnote #41) the strong relationship between inclusion in a primary disability (PRIMDIS) category and enrollment at a particular type of WPTI site. 60 This was not a chance association. It reflected both the secondary eligibility rules for WPTI, the tendency of community agencies to do outreach to current or former consumers, and in some cases state and/or organizational rules governing whom an agency could serve. “Agency type,” rather than “primary disability,” was emphasized in the modeling for the WPTI evaluation. Regression models restricted to WPTI participants (i.e. excluding members of the comparison group) exhibited the same general pattern of results as for the models incorporating PRIMDIS. 61 We believe that the evidence points to the dominant effect of agency recruitment, mission, and service philosophy as one principal cause of observed results. The cell sizes for combinations of PRIMDIS and WPTI agency type that do not “coincide” are modest, though it would make sense to at least examine their crosstabulation. 62 58 Delin, Barry S., Reither, Anne E., and Drew, Julia A. 2003. “Employment and Earnings Trends in a Complex Program Delivery Environment.” Presentation at the State Partnership Annual Meeting, Washington D.C., August 7-8, 2003. Average UI Earnings for employed participants at physical disability agencies in the eighth quarter following the enrollment quarter was $2,486. The amounts for those employed at mental health and developmental disability agencies were, respectively, $1,850 and $1,641. These data are for persons enrolled by March 31, 2001, six months prior to the final enrollment date used in the analyses contained in this paper. 59 See Delin, Barry S., et al. 2004. pp. 181-85 and p. 52. Unfortunately, no models examined the impact of PRIMDIS between WPTI participants and comparison group members. This was a result of using the PRIMDIS variable to compute the propensity scores needed to “equate” the two groups. 60 87% of those classified as “physical/HIV” in PRIMDIS received WPTI services at a designated “Physical Disability” or “HIV/AIDS” agency. The proportion of those with PRIMDIS = “affective” enrolled through “Mental Health” agencies was 72%.The proportion of those with PRIMDIS = “cognitive” enrolled through “Developmental Disability” agencies was 77%. See Delin, Barry S., et al. 2004. p. 79. 61 62 Delin, Barry S., et al. 2004. pp. 134-37. The data set used for this study does not include WPTI agency type, though that data can be obtained. Actual cells sizes are unknown. However, cell sizes range from a low of 1 to a high of 44 for the enrollment group of 527 cases from which the 506 were drawn. DRAFT – Do Not Cite Without Permission 30 Additionally, one of the two types of organizations, Community Support Programs, selected as WPTI “mental health” agencies used an agency wide service approach that encouraged employment mainly as a form of therapy and only secondarily as a means to increase economic welfare. These agencies, despite achieving high employment rates for their WPTI participants and, thereby, relatively high average earnings, were rated poorly on the implementation quality variable. 63 While it is doubtful that this alone accounts for the weak impact of the implementation quality (IMPQUAL) variable, it appears to have contributed to that result. Furthermore, we speculate that the combination of secondary eligibility criteria for the four types of WPTI provider agencies (physical disabilities, mental health, developmental disabilities, and HIV/AIDS), agency recruitment patterns, and agency service philosophies may explain why educational attainment and other sociodemographic variables appear to have little or no impact on whether a participant will have successful outcomes as we’ve defined them. This may also be the primary reason why attitudinal factors seem to have little effect either. This combination may be so powerful, in the specific context of how WPTI operated, as to swamp the expected effects of characteristics generally associated with labor market outcomes, especially those characteristics that might explain the different patterns in earnings and employment rates observed in the quarters prior to WPTI entry. We are, however, unwilling to speculate how frequently this type of selection bias might occur. While apparently not “creaming” in intent, it raises similar concerns about the ability to generalize results. 64 63 Based on our observations and interviews during WPTI, we think this observation would apply, if less stringently, to at least one of the “developmental disability” agencies. 64 It is important to note that random assignment of participants to treatment and control groups would not by itself have solved this problem. The problem is a consequence of the delivery system chosen for WPTI. That choice was not made for research considerations, but to exploit existing service delivery capacity and to work within an existing programmatic and legal framework. There is circumstantial evidence (mainly from interviews, surveys, and informal conversation) that limited “creaming” occurred. As service provision was, with few exceptions, provided through DVR authorizations (though the federal match was provided through DHFS funds), it is understandable that DVR counselors and community agency staff wanting to maintain good relationships with those counselors might give preference to individuals who appeared likely to achieve successful DVR closures (“twenty-sixes”). Though DVR state-level managers do not appear to have ever directly pressured counselors to withhold a DVR authorization for WPTI services for individuals who appeared unlikely to succeed, these managers did communicate their concern about the higher average cost per closure associated with WPTI compared to other patterns of DVR funded services. There is also some countervailing evidence that particular DVR counselors steered some particularly difficult to serve consumers to WPTI. We have no means of comparing the relative power of these two opposing effects on who entered the demonstration. However, it must also be remembered that the WPTI enrollment process incorporated “ability to benefit” judgments, making the identification of deliberate selection bias difficult. The first and most important was that made by an individual in deciding to enroll. From the perspective of DRAFT – Do Not Cite Without Permission 31 The issue is not, in our view, how the characteristics of WPTI participants differ from those of the working age SSDI and SSI (for reason of disability) population. It is how the characteristics of WPTI participants might differ from those within the population of the putatively eligible who might choose to utilize a program similar to WPTI should it become available through most, if not all, community agencies that provide services to persons with severe disabilities. The observed differences between WPTI participants and the comparison group of DVR consumers in their respective distributions for disability characteristics and some socio-demographic variables suggest that the characteristics of WPTI participants might not closely model those who might enter a similar program that was more widely available. In this vein, it is important to consider that a disproportionate amount of WPTI capacity was located at organizations that predominately served persons with physical disabilities. Should it be determined that the overall distribution of organizational capacity in Wisconsin is reasonably similar to that mobilized for the WPTI demonstration, we would have even greater confidence in using our results to inform policy design in at least that state. 65 Lastly, we noted that we had not included any of the baseline survey data about participants’ attitudes and perceptions in our models, largely because of the absence of association with our dependent variables at the .2 significance level and the large number of missing responses for certain items. Yet, in the WPTI evaluation, several of these variables proved to be significant predictors of employment status across time. 66 It seems counterintuitive that none of the attitudinal variables are associated with inclusion in the outcome groups. Thus, this is an area for further investigation, though with one exception we have not identified an approach to guide future work. The barriers measure included in the data set we used is an average score across a range of conceptually distinct impediments to employment or increasing work effort or earnings. Items include the participant’s perceptions of health status, needs to maintain access to health care and other public benefits, the availability of transportation and personal support, the respondents need for further education or training, and the impact of disincentives to earnings in Social Security and other public programs. Over the year between the baseline and follow-up surveys, WPTI participation appeared to WPTI’s designers and most of its organized stakeholders, the principle of consumer choice was paramount (among other things contributing to the decision not to use random assignment). Later Moreover, both the community agency and DVR counselor applied ability to benefit considerations. In the community agency’s case, this was supposed to be limited to having the ability to serve the individual’s specific needs, not just those of the typical individual fitting into the disability category the agency was authorized (for WPTI) to serve. The DVR counselor was to use the same criteria used for all other consumers. 65 WPTI did not formally structure the community agency selection process to result in having more of its enrollment capacity at the agencies serving those with “physical” disabilities. In the site selection RFP, WPTI offered that it was seeking a equal number of “physical disability” and “mental health” sites. In point of fact, the targeted enrollment capacity for each disability category that emerged from the agency selection process did not closely match the estimates of enrollment in each disability category that DHFS communicated to SSA, as part of Wisconsin’s SPI proposal. 66 These models contrasted all those in the WPTI study, not just those who received the intervention. See Delin, Barry S., et al. 2004. pp. 152 - 158. DRAFT – Do Not Cite Without Permission 32 lessen the perceived severity of barriers to employment in two ways, both through the intervention itself (e.g. benefits counseling and easy access to agency staff) and through the “reality testing” of achieving better employment outcomes. 67 It may be useful to explore changes in specific barrier items, treating the dependent indicator variables used in this paper (UI$I, UIEI, and FEI) as independent variables. To sum up, it appears certain that the differences between the more and less successful WPTI participants were strongly rooted in attributes and circumstances in place before entry into WPTI. The problem is that with the exception of maintaining attachment to work after entry to a Social Security program we have little to say what these were and, thus, have limited advice for those who might want to look for ways to target recruitment for return-to-work programs. We have found evidence to support that policy innovations like the Medicaid Buy-in and the SSI waiver can make a difference for those participants “poised” to make stronger (by our criteria) return-to-work efforts, but found no direct quantitative evidence of the role service provision aspects of WPTI, like benefits counseling, may play in facilitating better outcomes. Certainly, we will explore whether there are unexploited opportunities to look at this issue using available data. Finally, we are left with the unresolved issue of how typical WPTI participants are of the larger population that might choose to enter return-to-work programs and/or use existing or future work incentives. The WPTI experience suggests that organizational (rather than individual) selection can have a strong influence on individual participation and, it appears, outcomes. To the extent that delivery of employment and, more generally, social services continues to be highly decentralized in the United States, it is likely that such organizational factors matter. We cannot say, however, that they will matter in the same ways they did for WPTI. 67 Reither, Anne E., Delin, Barry S., and Drew, Julia A. 2005. WPTI Final Project Report – Addendum: Barriers to Employment. Menomonie, WI: University of Wisconsin – Stout Vocational Rehabilitation Institute. In particular, see pp. 10-12. DRAFT – Do Not Cite Without Permission 33 Appendix A: Independent and Control Variables Used in Logistic Models NAME AGE SEX RACE2 EDUCrev PRIMDIS OOSrev SSACat EMPRAT IMPQUAL BCtoQ5 VStoQ5 DESCRIPTION Age at Enrollment. Calculated in whole numbers at birthday on or preceding enrollment date. Sex. 1=male 2=female Race, recoded into a dichotomous variable. 1=other 2=white Education, recoded into three categories. 1=less than high school 2=high school diploma or GED 3=more than high school Primary Disability, coded into three categories. 1= cognitive/developmental 2=affective/mental 4=physical (including HIV/AIDS) Order of Selection assignment recoded. 1= most significant 2= other SSA status at enrollment. 1=SSDI or concurrent SSDI/SSI 2=SSI only. As the SSA administrative data available did not unequivocally identify this status on the enrollment date, any record of SSDI participation in the year leading up to enrollment resulted in that participant being assigned to the SSDI or concurrent . Ratio of years with reported earnings between entry to a Social Security disability program and WPTI enrollment (noninclusive). Values can range from “0” (no years with earnings) to “1” (all years having earnings). Average assessment of WPTI implementation quality as assessed by WPTI central program staff housed at DHFS and DVR. This variable measures perceived overall implementation quality, not that of the quality of the intervention received by any specific participant. 0=low 1=medium 2=high Hours of benefits counseling services delivered from enrollment quarter through the fifth quarter following enrollment. Hours of vocational services, including barriers assessment and vocational service planning, delivered from enrollment quarter through the fifth quarter following enrollment. These hours do not include services provided by vendors external to the WPTI agency. DATA SOURCE Encounter data Encounter data Encounter data Participant survey, supplemented with encounter data DVR administrative records, participant survey when RSA911 code unavailable. DVR administrative records SSA administrative records SSA administrative records and encounter data Q sort procedure Encounter data Encounter data DRAFT – Do Not Cite Without Permission JCtoQ5 SSIWAV MAPP Q0BS$_sdz FSQ0 Chg_sdz Hours of job coaching services delivered from enrollment quarter through the fifth quarter following enrollment. SSI Waiver participation (registering to use the waiver, if eligible) 0=never on waiver 1= on waiver at some time during WPTI Medicaid Buy-in participation 0=never on MAPP during WPTI 1= on MAPP at some time during WPTI Standardized amount of cash benefit (in 1996 GDP dollars) from SSA and the Wisconsin SSI state supplement during the enrollment quarter. Values were standardized by subtracting the mean and then dividing by the standard deviation. Food Stamp participation during the enrollment quarter 0=no 1=yes Standardized change in income proxy (in 1996 GDP dollars) between base (Q-1 through Q2) year and outcome year (Q5 through Q8). Income includes UI earnings, benefit payments from SSA and payments from the Wisconsin SSI state supplement. Values were standardized by subtracting the mean and then dividing by the standard deviation. 34 Encounter data WPTI administrative data DHFS administrative data Computed from SSA administrative data DWD administrative data Computed from SSA, DHFS, and DWD/UI administrative data DRAFT – Do Not Cite Without Permission 35 Appendix B: Frequencies for Categorical Variables by Success Indicators (UI$I, UIEI, & FEI) Categorical Variable Frequencies and Percentages for UI Earnings Success Indicator (UI$I) by “More Successful” and “Less Successful” Variable “More Successful” “Less Successful” SEX Male 142 (60%) 165 (61%) Female 94 (40%) 105 (39%) RACE2 White 190 (81%) 201 (74%) Other 46 (20%) 69 (26%) EDUCrev Less than H.S. 40 (17%) 52 (19%) H.S. 77 (33%) 74 (28%) More than H.S. 117 (50%) 141 (53%) PRIMDIS Cognitive/developmental 42 (18%) 53 (20%) Affective/mental 71 (30%) 39 (14%) Physical 123 (52%) 161 (66%) OOSrev Most significant 105 (46%) 149 (58%) Other 121 (54%) 110 (42%) SSACat SSDI or concurrent 177 (75%) 187 (69%) SSI only 59 (25%) 73 (31%) IMPQUAL Low 41 (17%) 39 (14%) Medium 101 (43%) 124 (46%) High 94 (40%) 107 (40%) SSIWAV Never 147 (62%) 214 (79%) At some time 89 (38%) 56 (21%) MAPP Never 180 (76%) 246 (91%) At some time 56 (24%) 24 (9%) FSQ0 No 202 (86%) 204 (76%) Yes 34 (14%) 66 (24%) Note: Percentages calculated for non-missing cases. There are 21 missing cases for OOSrev, 5 for EDUCrev. DRAFT – Do Not Cite Without Permission 36 Categorical Variable Frequencies and Percentages for UI Employment Success Indicator (UIEI) by “More Successful” and “Less Successful” Variable “More Successful” “Less Successful” SEX Male 106 (61%) 201 (61%) Female 68 (39%) 131 (40%) RACE2 White 148 (85%) 243 (73%) Other 26 (15%) 89 (27%) EDUCrev Less than H.S. 29 (17%) 63 (19%) H.S. 55 (32%) 96 (29%) More than H.S. 88 (51%) 170 (52%) PRIMDIS Cognitive/developmental 37 (21%) 58 (18%) Affective/mental 53 (31%) 57 (17%) Physical 84 (48%) 217 (65%) OOSrev Most significant 83 (50%) 171 (54%) Other 83 (50%) 148 (46%) SSACat SSDI or concurrent 132 (76%) 232 (70%) SSI only 42 (24%) 100 (30%) IMPQUAL Low 35 (20%) 45 (14%) Medium 66 (38%) 159 (48%) High 73 (42%) 128 (39)% SSIWAV Never 110 (63%) 251 (76%) At some time 64 (37%) 81 (24%) MAPP Never 126 (72%) 300 (90%) At some time 48 (28%) 32 (10%) FSQ0 No 149 (86%) 257 (77%) Yes 25 (14%) 75 (23%) Note: Percentages calculated for non-missing cases. There are 21 missing cases for OOSrev, 5 for EDUCrev. DRAFT – Do Not Cite Without Permission 37 Categorical Variable Frequencies and Percentages for “Forms” Reported Employment Success Indicator (FEI) by “More Successful” and “Less Successful” Variable “More Successful” “Less Successful” SEX Male 136 (60%) 171 (61%) Female 90 (40%) 109 (39%) RACE2 White 185 (82%) 206 (74%) Other 41 (18%) 74 (26%) EDUCrev Less than H.S. 36 (16%) 56 (20%) H.S. 73 (33%) 78 (28%) More than H.S. 114 (51%) 144 (52%) PRIMDIS Cognitive/developmental 53 (24%) 42 (15%) Affective/mental 59 (26%) 51 (18%) Physical 114 (50%) 187 (67%) OOSrev Most significant 105 (49%) 149 (55%) Other 109 (51%) 122 (45%) SSACat SSDI or concurrent 166 (74%) 198 (71%) SSI only 60 (27%) 82 (39%) IMPQUAL Low 36 (16%) 44 (16%) Medium 87 (39%) 138 (49%) High 103 (46%) 98 (35%) SSIWAV Never 148 (66%) 213 (76%) At some time 78 (35%) 67 (24%) MAPP Never 170 (75%) 256 (91%) At some time 56 (25%) 24 (9%) FSQ0 No 191 (85%) 215 (77%) Yes 35 (16%) 65 (23%) Note: Percentages calculated for non-missing cases. There are 21 missing cases for OOSrev, 5 for EDUCrev. DRAFT – Do Not Cite Without Permission 38 Appendix C: Selected Statistics for Interval Variables by Success Indicators (UI$I, UIEI, & FEI) Mean, Median and Standard Deviation for UI Earnings Success Indicator (UI$I) by “More Successful” and “Less Successful” Variable Mean Median Standard Approx. Deviation Mean/StDev. AGE “More Successful” 37.4 38.0 10.4 3.6 “Less Successful” 38.8 39.0 10.1 3.8 EMPRAT “More Successful” 0.66 1.00 .460 1.4 “Less Successful” 0.45 0.27 .457 1.0 BCtoQ5 “More Successful” 33.5 23.0 39.3 0.8 “Less Successful” 27.3 21.0 24.5 1.1 VStoQ5 “More Successful” 71.5 49.5 68.3 1.0 “Less Successful” 43.8 38.0 30.6 1.4 JCtoQ5 “More Successful” 4.6 0.0 17.0 0.3 “Less Successful” 1.1 0.0 5.4 0.2 SSA Q0 Benefit Amount (1996 GDP$)* “More Successful” $1758 $1735 $851 1.9 “Less Successful” $1983 $1928 $787 2.5 “Income” Change Base to Outcome Year (1996 GDP$)* “More Successful” $1525 $1031 $5925 0.3 “Less Successful” -$327 $50 $2041 0.2 * Presented statistics based on actual data rather than for standardized versions of these variables used in the regression models. DRAFT – Do Not Cite Without Permission 39 Mean, Median and Standard Deviation for UI Employment Success Indicator (UIEI) by “More Successful” and “Less Successful” Variable Mean Median Standard Approx. Deviation Mean/StDev. AGE “More Successful” 36.9 37.0 10.0 3.7 “Less Successful” 38.8 39.0 10.3 3.8 EMPRAT “More Successful” 0.65 1.00 .465 1.4 “Less Successful” 0.50 0.50 .464 1.1 BCtoQ5 “More Successful” 34.6 22.0 42.6 0.8 “Less Successful” 27.9 21.0 25.2 1.1 VStoQ5 “More Successful” 73.7 49.0 71.5 1.0 “Less Successful” 74.7 60.1 55.4 1.3 JCtoQ5 “More Successful” 5.2 0.0 18.9 0.3 “Less Successful” 1.5 0.0 6.5 0.2 SSA Q0 Benefit Amount (1996 GDP$)* “More Successful” $1733 $1739 $865 2.0 “Less Successful” $1953 $1893 $793 2.5 “Income” Change Base to Outcome Year (1996 GDP$)* “More Successful” $2251 $1100 $5963 0.4 “Less Successful” -$361 $59 $2945 0.1 * Presented statistics based on actual data rather than for standardized versions of these variables used in the regression models. DRAFT – Do Not Cite Without Permission 40 Mean, Median and Standard Deviation for “Forms” Reported Employment Success Indicator (FEI) by “More Successful” and “Less Successful” Variable Mean Median Standard Approx. Deviation Mean/StDev. AGE “More Successful” 38.1 38.3 9.9 3.9 “Less Successful” 38.2 39.0 10.5 3.6 EMPRAT “More Successful” 0.65 1.00 .457 1.4 “Less Successful” 0.47 0.33 .466 1.0 BCtoQ5 “More Successful” 34.7 22.5 39.3 0.9 “Less Successful” 26.6 21.0 25.1 1.1 VStoQ5 “More Successful” 77.9 54.5 71.6 1.1 “Less Successful” 71.4 59.5 51.5 1.4 JCtoQ5 “More Successful” 4.3 0.0 14.9 0.3 “Less Successful” 1.5 0.0 9.7 0.2 SSA Q0 Benefit Amount (1996 GDP$)* “More Successful” $1781 $1739 $805 2.2 “Less Successful” $1956 $1900 $832 2.4 “Income” Change Base to Outcome Year (1996 GDP$)* “More Successful” $1230 $384 $5031 0.2 “Less Successful” -$21 $70 $3743 0.1 * Presented statistics based on actual data rather than for standardized versions of these variables used in the regression models.